Transcribe every caller on hardware you own
Dograh is the self-hostable voice agent platform with an open-source workflow builder. Run Whisper or another open speech model on your own GPU through self-hosted Dograh, so no caller's voice travels to a speech provider.
How one sentence reaches your model
Run the model behind one route
/v1/audio/transcriptionsthe OpenAI-style transcription address
Speaches runs Whisper through faster-whisper. Parakeet or Qwen3-ASR need their own server that answers the same route.
→ Ring it yourself before callers do
Point the agent at your server
STT tabwhere a model configuration picks its speech model
Add a Local Models (Speaches) provider with your server's address, then select it on the STT tab.
→ Swap the English-only default for other languages
Watch what each turn costs
View Tracetiming for every turn of a call
Each pause sends the finished stretch of speech as one request, so the time on your card adds straight to the reply.
→ Compare models on the same set of calls
Pick by your callers' languages first
Parakeet ships under CC-BY-4.0, so NVIDIA gets credited wherever you ship it.
Quick to start with no hardware, though the provider receives every caller's audio.
Moving between columns changes a model name and an address, never the agent.
Prove the model on your own calls
- Real calls recordedTwenty to thirty of your own calls, with real customer names and order numbers, run through the server you will use.
- Phone audio heardMost calls arrive at 8 kHz, and some models that score well on clean recordings slip badly on a line.
- Silence checkedPlay silence and hold music through Whisper turbo, which can write out words nobody said.
- Credit line addedCC-BY-4.0 models generally need NVIDIA named on your notices page before you ship.
- Peak load triedSentences queue on a shared card, so load-test at the call count you expect at peak.
What owning the speech step gives you
Raw audio stays inside
With Dograh self-hosted and a local model, every caller's raw audio stays on servers you run.
A card, not a meter
Your GPU costs the same as call minutes climb. Power and engineer time still cost money.
Room left for the LLM
At roughly 2 GB, Parakeet TDT 0.6B v2 leaves most of a 24 GB card for the language model.
Hosted now, local later
Begin on a hosted provider and move the same agent to your GPU once call volume pays for the hardware.
What changes when speech runs in-house
| What you get | Dograh | Hosted speech-to-text APIs |
|---|---|---|
| Caller audio never leaves your network | ✓ | ✕ |
| No per-minute transcription charge | ✓ | ✕ |
| Open-source agent platform around the model (BSD-2) | ✓ | ✕ |
Choose your speech model, step by step
Which open speech model should I try first?+×
For English calls sharing a GPU with the language model, Parakeet TDT 0.6B v2. For many languages, try Whisper large-v3 or Qwen3-ASR.
How does Dograh send audio to my model?+×
At each pause, the finished segment goes as one request to your server's OpenAI-compatible /v1/audio/transcriptions route.
Does the default model understand Spanish or German?+×
No. Systran/faster-distil-whisper-small.en only knows English. Switch to Systran/faster-whisper-large-v3 and set the two-letter Language code.
Can I use Parakeet in a paid product?+×
Yes. Its CC-BY-4.0 licence generally allows paid use, provided NVIDIA is credited wherever the model ships.
How well does Parakeet hear phone audio?+×
Parakeet TDT 0.6B v2 averages a 6.32% word error rate after 8 kHz mu-law, against 6.05% on clean 16 kHz test sets.
How do I handle brand names a local model mishears?+×
Write the common mishearings into the prompt, and the language model corrects for them. The Dictionary setting serves hosted providers that take word lists.
Record your calls. Test two models on them.
Self-host in one command, or start on Dograh Cloud.
$ curl … start_docker.sh && ./start_docker.sh